Reinforcement Learning in Bioinformatics: Emerging Applications and Challenges

Reinforcement learning (RL) has moved from theoretical computer science into practical bioinformatics applications, particularly in drug discovery, protein design, and treatment optimization. This article explains how RL works in biological contexts, surveys current applications with evidence from peer-reviewed research, and provides a practical framework for evaluating whether RL is the right approach for a given bioinformatics problem. The framework helps students, researchers, and analysts decide when RL adds value over supervised learning or traditional computational methods, and how to design RL systems that produce biologically meaningful results.

What Reinforcement Learning Adds to Bioinformatics

Reinforcement learning differs fundamentally from supervised machine learning. In supervised learning, a model learns from labeled examples where the correct output is known. In RL, an agent learns by interacting with an environment, taking actions, and receiving rewards or penalties based on the outcomes of those actions. The agent's goal is to maximize cumulative reward over time through trial and error.

In bioinformatics, this paradigm fits problems where the solution space is vast and the path to a good solution is not obvious. Drug discovery provides a clear example. The chemical space of possible drug-like molecules is enormous, and searching it systematically is impractical. RL agents can learn to navigate this space by generating molecules, evaluating their properties, and using feedback to improve subsequent generations. This approach has been applied to molecular generation, peptide design, and protein engineering.

The key distinction for bioinformaticians is that RL excels at sequential decision-making problems. When a biological problem can be framed as a sequence of choices where each choice affects future options, RL becomes relevant. When the problem is simply predicting an outcome from static features, supervised learning remains more appropriate.

Core Principles of Reinforcement Learning in Biological Contexts

The Agent-Environment Framework

An RL system consists of an agent that takes actions and an environment that responds to those actions. In drug discovery, the agent might be a generative model that proposes molecular structures. The environment evaluates those structures using computational predictors for properties such as binding affinity, toxicity, or synthetic accessibility. The reward signal tells the agent how well its proposed molecules performed.

The environment in bioinformatics is almost always a computational surrogate instead of a physical laboratory. This distinction matters because computational predictors have their own error rates and limitations. An RL agent optimized against a flawed predictor will produce molecules that score well computationally but may fail in actual biological assays.

Reward Design as the Central Challenge

Reward design determines what the RL agent optimizes. In molecular generation, rewards might include predicted binding affinity, drug-likeness scores, synthetic accessibility estimates, or combinations of these factors. The choice of reward function shapes everything the agent learns.

Research on multi-property drug design demonstrates the complexity of reward design. A 2026 study developed a teacher-student-interaction deep learning model fine-tuned by reinforcement learning, called TSItransRL, to optimize multiple molecular properties simultaneously. The model achieved success rates of 98.36% for DRD2-targeting activity and 98.90% for dual JNK3/GSK3β-targeting activity, with an internal diversity of 0.795. The researchers used docking, similarity ensemble approach, scaffold novelty, and ADMET analyses as exploratory in silico assessments to support preliminary prioritization of generated molecules. Importantly, the authors noted that generated molecules should be regarded as computational candidates requiring further medicinal-chemistry assessment and independent validation.

This example illustrates two principles. First, multi-property optimization requires careful reward weighting. Second, computational success does not guarantee experimental success, so RL outputs must be treated as hypotheses instead of final answers.

Exploration Versus Exploitation

RL agents face a constant tradeoff between exploring new options and exploiting known good options. In biological sequence space, exploration means trying novel sequences that might have unexpected properties. Exploitation means refining sequences that already score well.

The exploration-exploitation balance is particularly important in bioinformatics because biological fitness landscapes are rugged. Small changes in sequence can produce large changes in function, and the space between good solutions may contain many poor solutions. Agents that exploit too heavily may get stuck in local optima. Agents that explore too much may waste computational resources on unpromising regions.

At a Glance: RL Suitability Assessment for Bioinformatics Problems

Applications in Drug Discovery and Molecular Design

Generative Models for Small Molecules

Reinforcement learning has become an integral part of modern ligand-based drug discovery. A 2020 review in Expert Opinion on Drug Discovery examined deep learning approaches to ligand-based novel drug discovery, covering virtual screening, neural generative models, and mutation-based structure generation. The authors considered multiple neural network architectures including deep multilayer networks, convolutional networks, recurrent networks, and autoencoders. They also examined different learning frameworks including adversarial learning and reinforcement learning, and different molecular representations including SMILES strings, graphs, and alternative string formats.

The review identified two persistent challenges for generative models in drug discovery: interpretability and explainability of deep learning models, and synthetic accessibility of novel compounds. These challenges remain relevant for RL-based approaches. A molecule that scores well computationally may be difficult or impossible to synthesize, limiting its practical value.

Deep Q-Networks for Molecular Generation

Deep Q-Networks, a classic RL architecture, have been applied to molecular generation. A 2025 IEEE conference paper titled "Accelerating Drug Discovery with Deep Reinforcement Learning: Molecular Generation Using Deep Q-Network" documents this approach. The method uses a deep neural network to estimate the expected cumulative reward of generating specific molecular structures, allowing the agent to select actions that lead to high-reward molecules.

The Q-network approach works well when the state space can be meaningfully represented and the reward signal is informative. For molecular generation, states might represent partial molecular structures and actions might represent adding specific atoms or bonds. The Q-network learns which structural additions lead to molecules with desirable properties.

Actor-Critic Methods for Complex Action Spaces

Drug discovery often involves multiple types of actions, such as choosing which part of a molecule to modify and then choosing how to modify it. Standard RL algorithms may struggle with these mixed action spaces. A 2021 AAAI conference paper introduced a "Towered Actor Critic" approach for handling multiple action types in reinforcement learning for drug discovery. The method uses separate network towers for different action types, allowing the agent to learn appropriate policies for each type of decision.

This architectural innovation addresses a practical problem in molecular design. Molecules have multiple modification points, and the optimal modification depends on the current state of the molecule. Towered architectures let the agent learn context-dependent policies for different action categories.

Geometric Deep Reinforcement Learning

Traditional molecular representations like SMILES strings struggle to capture three-dimensional spatial interactions. A 2026 study in the Journal of Computer-Aided Molecular Design introduced an integrated framework combining Geometric Multi-Discrete Soft Actor-Critic (Geom-SAC) with Multi-stage Variational Autoencoders (MS-VAE) for de novo molecular creation and activity optimization. The approach uses geometric deep learning to enforce physical atomic restrictions and a hierarchical VAE architecture to organize latent space into structural steps from scaffold formation to functional group optimization.

The method also incorporates a Non-Covalent Interaction-Aware graph neural network to improve protein-ligand affinity predictions by simulating intermolecular forces. On benchmark datasets including ZINC250k and PDBbind, the framework improved binding affinity scores by 15% and the Valid-Unique-Novel molecule ratio by 20% compared to existing methods. The study also described a blockchain-based security layer for data security and traceability.

This work demonstrates the value of incorporating physical constraints into RL systems. Molecules that violate physical laws are not viable drug candidates, so enforcing physical restrictions during generation reduces wasted computation.

Applications in Peptide and Protein Design

Reinforcement Learning for Peptide Generation

Peptides occupy an intermediate space between small molecules and full proteins. They offer advantages including precise control of amino acid sequence order, customization for property modification, and large-scale synthesis capability. A 2024 study in Briefings in Bioinformatics proposed a reinforcement learning-driven generation model based on graph attention mechanisms for peptide generation.

The model captures connectivity structures between amino acid residues in peptides using graph attention mechanisms. It uses an actor-critic framework with real-time feedback loops to achieve dynamic balance between attributes, allowing customization of peptide generation for specific targets and enhancement of peptide-target affinity. Experimental results showed that generated drug-like peptides met specified absorption, distribution, metabolism, excretion, and toxicity properties and bioactivity with a success rate over 90%.

The graph-based representation addresses a limitation of sequence-based approaches. Peptide function depends on three-dimensional structure, which depends on how amino acids interact spatially. Graph attention mechanisms capture these connectivity relationships more effectively than linear sequence representations.

Steering Generative Models for Protein Design

Generative AI models for protein design learn probability distributions from protein data and produce novel samples. However, because these models mirror their training distributions, they tend to sample from the most probable modes while low-probability regions, which often encode valuable properties, remain underexplored.

A 2026 review in Current Opinion in Structural Biology surveyed strategies for steering generative models toward user-specified properties. The review distinguished approaches that modify model parameters, such as reinforcement learning or supervised fine-tuning, from approaches that keep model parameters fixed, including conditional generation, retrieval-augmented strategies, Bayesian guidance, and tailored sampling methods.

Reinforcement learning fits into the parameter-modifying category. By providing rewards for proteins with desired properties, RL can push a generative model to produce proteins that differ from the training distribution in useful ways. This capability is valuable for designing proteins with novel functions or improved stability.

Antibody Design Platforms

Antibody therapeutics require accurate modeling of complex structures and interactions. A 2026 review in Topics in Current Chemistry examined AI-driven design platforms for next-generation antibody therapeutics. The review noted that emerging AI frameworks, including generative adversarial networks, reinforcement learning, and multi-omics integration, are accelerating target identification, optimizing lead candidates, and refining pharmacokinetic and biophysical profiles.

The review also discussed challenges in translating AI-driven strategies into clinical success. Antibodies are large, complex molecules, and their behavior in vivo depends on factors beyond binding affinity, including immunogenicity, half-life, and tissue penetration. RL systems that optimize only binding-related properties may produce antibodies that fail in later development stages.

Applications in Systems Pharmacology and Personalized Medicine

Moving Beyond Single-Target Drug Design

Many complex diseases, including neurological disorders, inflammatory diseases, and most cancers, are multi-genic systemic conditions that lack effective treatments. A 2022 review in Expert Opinion on Drug Discovery examined reinforcement learning powered systems pharmacology as a potential approach to designing personalized therapies for these diseases.

The review noted that while advanced RL techniques have been successfully applied to target-based drug discovery, new RL strategies are needed for systems pharmacology-oriented personalized de novo drug design. Single-target approaches may miss the complexity of diseases that involve multiple biological pathways and feedback loops.

Systems pharmacology requires RL agents to consider entire biological networks instead of individual molecular targets. This increases the dimensionality of the problem and complicates reward design. The reward function must capture therapeutic benefit across multiple pathways while avoiding unintended effects.

Treatment Planning as Sequential Decision Making

Treatment planning naturally fits the RL framework because it involves a sequence of decisions over time. A clinician decides which treatment to administer, observes the patient's response, and adjusts subsequent treatments accordingly. RL can learn optimal treatment policies from historical patient data.

However, treatment planning RL systems face significant challenges. Patient data are heterogeneous and incomplete. Clinical outcomes take time to observe, creating delayed reward signals. And the cost of exploration is high because poor treatment decisions can harm patients. These factors limit the practical deployment of RL for treatment planning, despite its theoretical appeal.

The Role of Data Resources and Infrastructure

Public Data Repositories

RL systems in bioinformatics depend on high-quality training data and evaluation resources. The European Bioinformatics Institute provides training and data resources for bioinformatics research, including databases for sequences, structures, and functional annotations. The National Center for Biotechnology Information maintains comprehensive data resources including GenBank, PubMed, and various molecular databases.

These resources serve multiple purposes in RL workflows. They provide training data for generative models, evaluation data for reward functions, and validation data for assessing whether RL-generated molecules resemble known bioactive compounds. Researchers should document which data resources they use and how those resources influence model behavior.

Data Sharing and FAIR Principles

The NIH Genomic Data Sharing Policy establishes expectations for sharing genomic data generated through NIH-funded research. The policy addresses data deposition, access, and responsible use. Researchers developing RL systems that use genomic data should understand and comply with applicable data sharing requirements.

The FAIR Guiding Principles, published in Scientific Data, provide a framework for making data findable, accessible, interoperable, and reusable. These principles apply to RL training data, evaluation datasets, and model outputs. Following FAIR principles improves reproducibility and enables other researchers to build on published work.

For RL systems, FAIR principles have specific implications. Training data should be described with sufficient metadata to understand its provenance and limitations. Evaluation protocols should be documented so others can replicate results. Generated molecules or sequences should be deposited in accessible repositories with clear descriptions of how they were produced.

Practical Framework for Evaluating RL Approaches

Step 1: Define the Problem Structure

Before adopting RL, determine whether the problem involves sequential decision making. RL is appropriate when the solution requires a sequence of choices where each choice affects future options. Problems that involve a single prediction from static features are better suited to supervised learning.

For drug design, the sequential structure is clear. A molecule is built atom by atom or fragment by fragment, and each addition affects which subsequent additions are possible or desirable. For protein design, the sequence of amino acid choices similarly has sequential structure.

Step 2: Assess the Reward Signal

RL requires a reward signal that accurately reflects the desired outcome. In bioinformatics, rewards are typically computational predictions instead of experimental measurements. Assess the reliability of available predictors before using them as rewards.

If binding affinity predictions have high error rates, an RL agent optimized against those predictions may produce molecules that score well computationally but fail experimentally. Consider whether multiple predictors can be combined to create a more robust reward signal.

Step 3: Evaluate Sample Efficiency Requirements

RL is notoriously sample-inefficient, requiring many interactions with the environment to learn good policies. In bioinformatics, each interaction involves generating a candidate molecule or sequence and evaluating it computationally. While computational evaluation is cheaper than experimental evaluation, it still consumes significant computing resources.

Consider whether the available computational budget supports the number of interactions needed for RL to converge. For problems where evaluation is expensive, alternative approaches such as Bayesian optimization may be more practical.

Step 4: Consider the Exploration Space

RL agents explore the solution space during training. In bioinformatics, the solution space is often vast. The chemical space of drug-like molecules is estimated to be enormous, and protein sequence space is similarly large.

Assess whether the exploration strategy can meaningfully cover the relevant regions of the solution space. If the space is too large for effective exploration, consider whether the problem can be decomposed into smaller subproblems or whether additional constraints can narrow the search.

Step 5: Plan for Validation

RL-generated molecules or sequences are computational hypotheses that require experimental validation. Plan validation experiments before starting RL training. Consider which experimental assays will be used, how many candidates can be tested, and what criteria will determine whether a candidate advances.

The 2026 TSItransRL study provides a model for this approach. The researchers used docking, similarity ensemble approach, scaffold novelty, and ADMET analyses as exploratory in silico assessments to support preliminary prioritization of selected generated molecules, while clearly stating that generated molecules should be regarded as computational candidates for further medicinal-chemistry assessment.

Common Failure Patterns in RL for Bioinformatics

Reward Hacking

Reward hacking occurs when an RL agent finds ways to maximize reward that do not actually achieve the intended goal. In drug design, an agent might generate molecules that score well on computational predictors but are not synthesizable or biologically active.

The 2020 ligand-based drug discovery review identified synthetic accessibility as a persistent challenge. RL agents optimized purely for predicted activity may generate molecules that cannot be synthesized. Including synthetic accessibility estimates in the reward function can mitigate this problem, but these estimates are themselves imperfect.

Overfitting to Training Distribution

Generative models trained on existing molecules tend to reproduce the properties of their training sets. RL can push models toward novel regions of chemical or sequence space, but excessive novelty may produce molecules that are unlike known bioactive compounds and therefore unlikely to be viable.

The protein design review noted that generative models mirror their training distributions and tend to sample from the most probable modes. RL steering can overcome this tendency, but the balance between novelty and drug-likeness requires careful reward design.

Computational Predictor Errors

RL agents optimize against computational predictors, and predictor errors propagate through the optimization process. If a toxicity predictor has false negative rates, the RL agent will learn to generate molecules that exploit those errors.

Researchers should characterize predictor error rates before using predictors as rewards. Cross-validation against held-out experimental data provides one approach. Where possible, use ensemble predictors or uncertainty-aware predictors to reduce the impact of individual model errors.

Evaluation Metric Mismatch

The metrics used to evaluate RL performance may not match the metrics that matter for real-world application. A model might achieve high scores on computational benchmarks while producing molecules that fail in biological assays.

The 2026 geometric deep reinforcement learning study reported improvements in binding affinity scores and Valid-Unique-Novel molecule ratios. These metrics measure computational performance, not experimental success. Researchers should be explicit about the distinction between computational benchmarks and biological outcomes.

Records and Measurements for RL Projects

Documentation Requirements

RL projects in bioinformatics generate large amounts of data that should be systematically recorded. Essential records include training data provenance, model architecture choices, hyperparameter settings, reward function definitions, and evaluation protocols.

The FAIR Guiding Principles provide a framework for organizing these records. Training data should be described with metadata that makes its provenance and limitations clear. Model configurations should be versioned and documented. Evaluation results should be reported with sufficient detail to enable replication.

Performance Metrics

Standard metrics for generative molecular design include validity, uniqueness, novelty, and property satisfaction rates. The TSItransRL study reported success rates for generating molecules with specific target activities, internal diversity, and scaffold novelty. The geometric deep learning study reported binding affinity improvements and Valid-Unique-Novel molecule ratios.

Choose metrics that align with the intended application. For drug discovery, property satisfaction rates matter more than raw novelty. For protein design, functional activity and stability matter more than sequence diversity.

Computational Resource Tracking

RL training consumes substantial computational resources. Track GPU hours, CPU time, memory usage, and storage requirements. This information supports cost estimation for scaling up projects and helps identify inefficiencies in the training pipeline.

Quality Controls and Validation Strategies

Computational Validation

Before experimental testing, RL-generated molecules should pass multiple computational filters. Standard filters include drug-likeness rules, synthetic accessibility scores, and predicted toxicity assessments. The 2026 TSItransRL study used docking, similarity ensemble approach, scaffold novelty, and ADMET analyses as exploratory in silico assessments.

These filters reduce the number of candidates requiring experimental testing, but they do not guarantee experimental success. Computational predictions have inherent uncertainty, and the filters themselves may have error rates.

Experimental Validation Planning

Plan experimental validation before starting RL training. Define the number of candidates that can be tested, the assays that will be used, and the criteria for advancing candidates to later development stages. The 2026 antibody design review noted that AI-driven strategies face challenges in translating to clinical success, emphasizing the importance of rigorous experimental validation.

Reproducibility Controls

RL systems involve stochastic elements, including random initialization and exploration noise. Reproducibility requires fixing random seeds, documenting data splits, and versioning code and data. The ACEGEN toolkit, described in a 2024 Journal of Chemical Information and Modeling paper, provides a streamlined framework for generative drug design built on TorchRL, with thoroughly tested reusable components. Tools like this support reproducibility by providing standardized implementations.

Limitations and Interpretation Boundaries

Computational Candidates Versus Experimental Results

RL-generated molecules and sequences are computational candidates, not validated therapeutics. The 2026 TSItransRL study explicitly stated that generated molecules should be regarded as computational candidates for further medicinal-chemistry assessment and independent validation. This distinction is critical for interpreting RL results.

Generalization Limits

RL models trained on specific data distributions may not generalize to related but different problems. A model trained on kinase inhibitors may not perform well on other target classes. Researchers should test generalization explicitly instead of assuming it.

Time and Cost Considerations

Drug discovery remains a costly and time-consuming process. A 2023 review in the International Journal of Molecular Sciences noted that advancing a new medication typically costs approximately USD 2.6 billion and consumes a mean time span of 12 years. AI and deep learning approaches, including RL, aim to reduce these costs, but they do not eliminate the fundamental challenges of drug development.

The Role of Human Expertise

RL systems augment instead of replace human expertise. The 2021 Molecular Diversity review noted that artificial intelligence and machine learning technology play a crucial role in drug discovery and development, with machine learning and deep learning algorithms implemented in processes including peptide synthesis, structure-based virtual screening, ligand-based virtual screening, toxicity prediction, drug monitoring and release, pharmacophore modeling, quantitative structure-activity relationship, drug repositioning, polypharmacology, and physiochemical activity. However, medicinal chemists and biologists must interpret RL outputs, design validation experiments, and make decisions about which candidates to advance.

Safety and Regulatory Context

Data Privacy and Security

RL systems in bioinformatics may use patient data, particularly for treatment planning and personalized medicine applications. The NIH Genomic Data Sharing Policy establishes expectations for responsible use of genomic data. Researchers must ensure compliance with applicable privacy and security requirements.

The 2026 geometric deep reinforcement learning study described a blockchain-based security layer for data security and traceability. While blockchain is not standard practice, the emphasis on data security reflects the importance of protecting sensitive biological and clinical data.

Responsible AI Use

RL systems that influence drug development or treatment decisions carry responsibility for patient safety. Computational predictions should not override clinical judgment or established safety protocols. The 2019 Neurochemistry International review of GLP-1 signaling illustrates the complexity of biological systems, where hormones like GLP-1 modulate multiple neural circuits and affect behavior beyond their peripheral actions. RL systems that model such systems must account for this complexity.

Escalation Criteria

Researchers should define criteria for escalating concerns about RL system behavior. Red flags include reward hacking, unexpected degradation of generated molecule quality, systematic biases in generated sequences, and discrepancies between computational predictions and experimental results. When these issues arise, the appropriate response is to pause deployment, investigate the cause, and adjust the system before continuing.

Frequently Asked Questions

What types of bioinformatics problems are best suited for reinforcement learning?

Problems with sequential decision-making structure are best suited for RL. Drug design fits this pattern because molecules are built through sequences of additions or modifications. Protein design similarly involves sequential choices of amino acids or structural elements. Problems that involve a single prediction from static features are better addressed with supervised learning. Treatment planning also fits RL because it involves sequences of decisions over time, though practical deployment faces significant challenges.

How does reinforcement learning differ from supervised learning in drug discovery?

Supervised learning trains models on labeled examples where the correct output is known, such as molecules with measured binding affinities. RL trains agents through interaction with an environment, where the agent proposes molecules and receives reward signals based on predicted properties. RL can explore regions of chemical space not represented in training data, while supervised learning is limited to interpolating between known examples. However, RL requires careful reward design and is computationally more expensive.

What are the main challenges in applying reinforcement learning to drug design?

The main challenges include reward design, sample efficiency, and validation. Reward functions must accurately capture the desired molecular properties, which requires reliable computational predictors. RL is sample-inefficient, requiring many evaluations to learn good policies. And RL-generated molecules are computational candidates that require experimental validation, which is costly and time-consuming. The 2020 ligand-based drug discovery review also identified interpretability and synthetic accessibility as persistent challenges.

How do researchers design reward functions for molecular generation?

Reward functions typically combine multiple property predictions, including binding affinity, drug-likeness, synthetic accessibility, and toxicity. The weights assigned to different properties determine what the RL agent optimizes. The 2026 TSItransRL study demonstrated multi-property optimization using bioactivity datasets and achieved high success rates for generating molecules with specific target activities. Reward design requires balancing competing objectives and understanding the error rates of the underlying predictors.

What is the difference between actor-critic methods and Q-learning for drug discovery?

Q-learning estimates the expected cumulative reward of taking specific actions in specific states, allowing the agent to select actions that lead to high-reward outcomes. Actor-critic methods maintain separate networks for the policy (actor) and value estimation (critic), which can handle more complex action spaces. The 2021 Towered Actor Critic paper addressed the challenge of multiple action types in drug discovery, where the agent must decide both which part of a molecule to modify and how to modify it.

How do geometric deep learning approaches improve reinforcement learning for drug design?

Geometric deep learning incorporates three-dimensional spatial information into molecular representations, addressing limitations of string-based representations like SMILES. The 2026 Geom-SAC study combined geometric deep learning with hierarchical variational autoencoders to enforce physical atomic restrictions and organize latent space into structural steps. This approach improved binding affinity scores by 15% and Valid-Unique-Novel molecule ratios by 20% on benchmark datasets compared to existing methods.

What validation is required before RL-generated molecules can be considered for drug development?

RL-generated molecules require both computational and experimental validation. Computational validation includes docking studies, similarity assessments, scaffold novelty analysis, and ADMET predictions. Experimental validation involves synthesizing candidate molecules and testing them in biological assays. The 2026 TSItransRL study used computational assessments for preliminary prioritization while clearly stating that generated molecules should be regarded as computational candidates for further medicinal-chemistry assessment and independent validation.

How can researchers evaluate whether reinforcement learning is the right approach for their bioinformatics problem?

Researchers should assess whether the problem has sequential decision-making structure, whether reliable reward signals are available, whether the computational budget supports RL training, and whether validation pathways exist. If the problem involves a single prediction from static features, supervised learning is more appropriate. If the problem involves sequential choices with feedback, RL may add value. The framework in this article provides a structured approach to this evaluation.

Related Bioinformatics Guides

• Deep Learning for Protein-Ligand Binding Affinity Prediction in Antiviral Drug Design
• Deep Learning in Protein-Ligand Binding Affinity Prediction for Antiviral Drug Design
• Network Pharmacology Approaches to Multi-Target Drug Discovery
• Machine Learning Approaches in Structural Virology: Predicting Viral Protein Dynamics
• The Bioinformatics Revolution in Structural Proteomics and Computational Drug Discovery: A Unified Paradigm

References and Further Reading

• EMBL-EBI Training. European Bioinformatics Institute.
• NCBI Data Resources. National Center for Biotechnology Information.
• Genomic Data Sharing Policy. National Institutes of Health.
• The FAIR Guiding Principles. Scientific Data.
• Artificial intelligence to deep learning: machine intelligence approach for drug discovery.. Molecular diversity, 2021.
• Artificial Intelligence and Machine Learning Technology Driven Modern Drug Discovery and Development.. International journal of molecular sciences, 2023.
• Integrating QSAR modelling and deep learning in drug discovery: the emergence of deep QSAR.. Nature reviews. Drug discovery, 2024.
• Reinforcement learning for systems pharmacology-oriented and personalized drug design.. Expert opinion on drug discovery, 2022.
• ACEGEN: Reinforcement Learning of Generative Chemical Agents for Drug Discovery.. Journal of chemical information and modeling, 2024.
• GLP-1: Molecular mechanisms and outcomes of a complex signaling system.. Neurochemistry international, 2019.
• Reinforcement learning-driven exploration of peptide space: accelerating generation of drug-like peptides.. Briefings in bioinformatics, 2024.
• The power of deep learning to ligand-based novel drug discovery.. Expert opinion on drug discovery, 2020.
• Multi-Property De Novo Drug Design Using Deep Learning-Based Knowledge Distillation and Reinforcement Learning.. 2026.
• Steering generative models for protein design: Aligning and conditioning strategies.. 2026.
• AI-Driven Design Platforms of Next-Generation Antibody Therapeutics.. 2026.
• Geometric deep reinforcement learning with hierarchical variational autoencoders for de novo drug design and activity optimization.. 2026.
• Drug Discovery using Generative Adversarial Network with Reinforcement Learning. 2021 International Conference on Computer Communication and Informatics Iccci 2021, 2021.
• Accelerating Drug Discovery with Deep Reinforcement Learning: Molecular Generation Using Deep Q-Network. Proceedings 2025 IEEE International Conference on Information Reuse and Integration and Data Science Iri 2025, 2025.
• Towered Actor Critic For Handling Multiple Action Types In Reinforcement Learning For Drug Discovery. 35th Aaai Conference on Artificial Intelligence Aaai 2021, 2021.

This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.